Avoiding the AI look: a quality check for client-ready AI video

Most ranking pages on this topic say "look for weird hands" and stop. Agency work needs more: a list your whole team uses, a rule for when a clip is rejected, a way to decide whether to fix or regenerate, and a record of who approved it. This page gives you those. It is tool-neutral and works on clips from any generator.

The cut-offs below are my proposal for a small agency, built on the published guides I list in the sources. They are not an industry standard. Change them for your clients.

What "the AI look" actually is

It is not one thing. Morphic's glossary separates artifacts (visible faults) into four types: compression artifacts from file handling, generation artifacts such as bad hands, merged faces and invented text, temporal artifacts such as flicker and ghosting between frames, and semantic artifacts where the picture looks fine but is wrong, for example a label that says something else. The first is a file problem. The other three are the AI look.

Two cautions. First, visual tells are unreliable as evidence. Magic Hour's own guide says visual anomalies and detector scores support investigation but are not proof, and a missing content credential does not show a clip is real or fake. So do not tell a client the work is "undetectable". Say it is "reviewed, with no known faults, and labelled where required". Second, a clean-looking clip can still fail on facts.

The five fault families and what each looks like

1. Bodies and identity. Extra or missing fingers, hands that merge with objects, a face that drifts between the first and last frame, teeth that change.
2. Text and brand. Garbled lettering on signs, packs and screens, a logo with the wrong shape, brand colour pushed to a different shade or over-saturated.
3. Physics and motion. Liquid that flows the wrong way, objects that pass through each other, flickering edges, motion that is too smooth, odd walking.
4. Light and look. Lighting that changes between cuts, flat light with no source, reflections that do not match, textures that repeat.
5. Sound and speech. Lip movement that misses the sounds b, p and m, a flat computer voice, no breath, background audio that jumps.

Some of these also appear in real edited footage, so judge the whole clip against the brief rather than hunting for a single tell.

The 12-point review checklist

Items 1 to 4 are hard fails: any one means reject. Items 5 to 12 are marked fix or ship.

#	Check	Pass looks like	If it fails
1	Fingers and hands at every pause	Correct count and shape in every held frame	Hard fail: regenerate, or reframe so hands are out of shot
2	Any text, logo or brand name	Letters and mark identical to the approved file	Hard fail: remove it and place the real one in an editor
3	Identity, first frame against last frame	Same person, same product, same clothes	Hard fail: regenerate with a reference image
4	Every claim, number and price	Appears in the client's approved script or sheet	Hard fail: correct the script and redo or re-edit
5	Brand colour	Primary colour on product and logo within the tolerance agreed with the client	Fix: grade in post
6	Motion and physics	Natural movement, no flicker or ghosting	Cut around it if 1 to 2 seconds, else regenerate
7	Lighting continuity	One light direction across cuts	Fix with a grade, or regenerate the odd clip
8	Lip sync and speech	Mouth matches the sounds	Regenerate lipsync, or cover with a cutaway
9	Audio	Clean, consistent volume, no jumps	Fix in the audio edit
10	Edges, backgrounds and small objects	No melting, stray objects or tile patterns	Retouch, or crop
11	Platform format	Ratio, safe zones, captions, file size, first-frame hook	Re-export or re-edit
12	People and likeness	Every real person consented; no one resembles a real person without permission	Remove the clip. Do not fix this by editing.

Pass/fail beats a score. Published guides such as Apiframe's give a 1 to 5 scale for prompt match, visual quality, motion, consistency and audio sync, but set no cut-off. A score invites arguing about whether a 3 is good enough. A hard-fail list does not.

How to review: five passes in a fixed order

Each pass has one purpose. Doing them in a set order catches more than watching the clip once.

1. Sound off, full screen. Judge the picture alone.
2. Frame by frame at hands, mouth, text and edges. Step through, or pause every second.
3. First frame against last frame. Look for identity drift in faces and products.
4. Sound on. Lip sync, voice, music, jumps.
5. Phone, normal speed. How it will actually be seen, with the platform's captions and crop.

Apiframe suggests at least two reviewers who score important clips independently, without knowing which model made which clip. In a small agency that means one person who did not write the prompt. People who prompted a clip tend to see what they meant, not what is on screen.

Regenerate, fix in post, or cut around it

This rule is a synthesis of the Morphic, Genra and Apiframe guides, so treat it as a working rule:

• Fix by hand when the fault is isolated in space: text, a logo, a colour, a caption, a stray object. Text and logos should be added in post anyway.
• Regenerate when the fault is structural or runs through time: anatomy, identity drift, flicker across the clip, physics. Morphic advises fixing the input (a cleaner reference or a simpler pose) and generating again for warped anatomy and drifting identity. Inpainting (repainting a small area), upscaling and compositing are the usual tools for smaller faults.
• Cut around it when a bad span is 1 to 2 seconds and the edit still works without it.

Before you regenerate, change one thing. Morphic's troubleshooting guide says "Change one thing first, then regenerate." If you change five, you pay for new faults and do not learn what fixed the old ones. Genra recommends 3 to 5 second segments stitched together, because shorter clips drift less and are cheaper to redo. For softness, an upscaler may be cheaper than a new generation. See upscaling AI video and editing tools in Higgsfield.

What never to put in a prompt

Brand names, prices, legal lines, phone numbers and web addresses are placed from approved files in an editor. Never type them into a prompt and hope. For images, see getting correct text into AI images. Keep faces of real people out unless you have written consent. See the rights checklist for AI video.

Similarly, keep hands out of frame where you can. Higgsfield's own hands and faces page advises training a Soul ID (a saved face model) from 20 or more reference photos and describing where the hand starts and ends. It shows no success rates, so I would not treat it as a cure, and I do not claim any tool handles hands best.

What a reject costs: retry-rate math

Quote clients for keepers, not generations. Here is an illustration with made-up numbers. You need ten 5-second clips. If 1 in 4 clips passes, you make 40. If each generation costs $1 in credits, that is $40. If a reviewer needs 15 minutes per 4 clips, review time is 2.5 hours, which at $45 an hour is about $112. The review costs more than the credits. One vendor, invideo, says on its own FAQ pages that it discarded about 85% of the clips in one localization project, and it puts the share of clips used in other productions much higher. That is a single vendor-reported outlier, so use it as a reason to measure your own rate, not as a benchmark.

Track two numbers per job: clips generated per keeper, and review minutes per keeper. If either doubles, change the process (shorter clips, better reference images) before you change the price. For pricing, see how to price AI-made creative.

Who signs off and what you record

One named person signs off every client deliverable. For each approved output, record:

• Tool and model name, and version.
• The prompt, seed (if the tool shows one) and any reference files.
• Who reviewed it, which checklist items were marked, and the date.

This helps with disclosure and rights questions later, and it lets you recreate or fix the clip. Sign-off mechanics are in the creative approval workflow, and a shared record in an AI asset library.

What to tell the client

Say what you checked: "Reviewed against a 12-point list, no known faults, text and logos placed from your approved files." Do not say "undetectable" and do not say "perfect". Whether to tell a client that AI was used, and whether the platform needs a label, depend on your contract, local law and the platform's current rules. Check each. This is information, not legal advice.

When to switch to live footage

• Close-ups of hands doing a precise task are the whole point of the video.
• A real person's testimony or a real result is the claim.
• The product has fine print or reflective surfaces that you cannot afford to get wrong.
• Your reject rate stays above what the budget can bear after two rounds of process changes.

Also compare cost with stock or filming in AI video vs stock footage vs filming yourself. A human editor or compositor is also the better tool for text, logos and brand colours. Regenerating for those is wasted money.

How I'd approach this

I would print the 12 points, make the first four hard fails and ask someone who did not prompt the clip to run the five passes. I would log every reject with its cause for a month, then look at the pattern. Repeated causes tell you what to put in your SOP and which prompts to retire.

Blur Studio builds the repeatable side of this for agencies: a review checklist inside an approval portal, a shared asset library and a record of which tool and prompt made each file. See the business page for how it works.